Renewable-Energy-Aware Predictive Computing for Generative AI: Forecasting Solar and Wind Energy Availability for Carbon-Aware Computational Scheduling

The increasing computational demands of generative artificial intelligence (AI), particularly large language model (LLM) inference, have created growing energy requirements for modern AI infrastructure. Although existing computational optimization approaches investigate workload forecasting, off-peak precomputation, dynamic electricity pricing, and low-cost GPU resource allocation, these mechanisms do not necessarily determine whether sufficient renewable energy will be available when computational workloads are executed. Electricity prices and renewable energy availability represent related but fundamentally different operational signals. A period with relatively inexpensive electricity does not necessarily correspond to a period with abundant solar or wind generation, while renewable energy production may fluctuate substantially even when electricity prices remain relatively stable. This paper introduces Renewable-Energy-Aware Predictive Computing for Generative AI (REAPC-GAI), a conceptual framework that integrates renewable energy availability forecasting with predictive computational preparation and adaptive inference scheduling. The proposed framework investigates how anticipated solar generation, wind generation, weather conditions, regional renewable energy production, energy storage availability, and grid-level energy composition can be translated into predictions of future computational energy availability. Its architecture incorporates six interconnected functions: renewable energy signal acquisition, probabilistic renewable generation forecasting, computational energy demand estimation, renewable-aware workload selection, adaptive computational scheduling, and realized environmental benefit evaluation. Rather than simply predicting future renewable electricity generation, REAPC-GAI connects energy supply forecasts to the estimated power consumption and temporal flexibility of individual computational workloads. A unified optimization formulation is developed to determine when predicted workloads should be prepared, how much computational capacity should be allocated, and whether tasks should be executed locally, deferred, or distributed across alternative locations according to forecast renewable energy availability. The formulation accounts for forecast uncertainty, GPU power consumption, renewable generation variability, grid electricity requirements, storage constraints, task deadlines, computational preparation costs, and service-level requirements. Particular attention is given to the distinction between renewable energy generation forecasts, physically accessible renewable electricity, contractual renewable energy procurement, and grid-average carbon intensity. An experimental methodology is proposed to evaluate renewable forecasting accuracy, renewable energy utilization, GPU energy consumption, carbon emissions, computational expenditure, workload completion time, and prediction-related operational risks. The framework does not assume that renewable-energy-aware scheduling automatically reduces emissions or costs; its effectiveness depends on renewable supply predictability, workload flexibility, electricity infrastructure, energy accounting boundaries, and the compatibility between forecast energy availability and computational demand. By connecting renewable energy forecasting with predictive inference preparation, this research investigates a potential transition from economically optimized predictive computing toward energy-supply-aware predictive computational orchestration for generative AI.